Jingxiu Yao

Heterogeneous defect prediction (HDP) is a significant research topic in cross-project defect prediction (CPDP), due to the inconsistency of metrics used between source and target projects. While most HDP methods aim to improve the performance of models trained on data from one source project, few studies have investigated how the number of source projects affects predictive performance. In this paper, we propose a new multi-source heterogeneous kernel mapping (MSHKM) algorithm to analyze the effects of different numbers of source projects on prediction results. First, we introduce two strategies based on MSHKM for multi-source HDP. To determine the impact of the number of source projects on the predictive performance of the model, we regularly vary the number of source projects in each strategy. Then, we compare the proposed MSHKM with state-of-the-art HDP methods and within-project defect prediction (WPDP) methods, in terms of three common performance measures, using 28 data sets from five widely used projects. Our results demonstrate that, (1) in the multi-source HDP scenario, strategy 2 outperforms strategy 1; (2) for MSHKM, a lower number of source projects leads to better results and performance under strategy 1, while n = 4 is the optimal number under strategy 2; (3) MSHKM performs better than related state-of-the-art HDP methods; and (4) MSHKM outperforms WPDP. In summary, our proposed MSHKM algorithm provides a promising solution for heterogeneous cross-project defect prediction, and our findings suggest that the number of source projects should be carefully selected to achieve optimal predictive performance.

Software defect prediction (SDP) is an effective technique to lower software module testing costs. However, the imbalanced distribution almost exists in all SDP datasets and restricts the accuracy of defect prediction. In order to balance the data distribution reasonably, we propose a novel resampling method LIMCR on the basis of Naïve Bayes to optimize and improve the SDP performance. The main idea of LIMCR is to remove less-informative majorities for rebalancing the data distribution after evaluating the degree of being informative for every sample from the majority class. We employ 29 SDP datasets from the PROMISE and NASA dataset and divide them into two parts, the small sample size (the amount of data is smaller than 1100) and the large sample size (larger than 1100). Then we conduct experiments by comparing the matching of classifiers and imbalance learning methods on small datasets and large datasets, respectively. The results show the effectiveness of LIMCR, and LIMCR+GNB performs better than other methods on small datasets while not brilliant on large datasets.